EU AI Act transparency duties apply now; high-risk duties from December 2027. Check your exposure
Insights About us Careers
Contact us
Retail & E-commerce

AI for Grocery Retailers

Grocery is retail with a clock attached. A forecasting error on a jumper becomes a markdown; the same error on strawberries becomes waste, and it happens every single day across every store.

Tier 2
Our depth here
Fresh
Where the value is
Daily
The forecast cadence

Fresh is where grocery AI earns its money, because it is the only category where being wrong in either direction costs you within forty-eight hours and where the volume of decisions is far beyond what any person can review.

In one paragraph

AI for grocery covers fresh and perishable demand forecasting, waste reduction and dynamic markdown, ambient replenishment, substitution recommendation for online orders, picking and fulfilment optimisation, and promotional planning.

Fresh forecasting is a different problem from ambient

Ambient grocery forecasting is ordinary retail forecasting. Fresh has a shelf life measured in days, a delivery cadence measured in hours, weather sensitivity that is real rather than folklore, and a cost of over-ordering that is total rather than deferred.

FactorAmbientFresh
Forecast horizonWeeksOne to three days
Cost of over-orderingCarrying cost, later markdownTotal loss within days
Cost of under-orderingLost saleLost sale plus a customer who shops elsewhere
Weather sensitivityModestVery high in some categories
Waste dataNot applicableThe most valuable series you have, and often unused
Markdown decisionSeasonalDaily, in-store, time-of-day
Worth knowing

Your waste data is the signal nobody models

Most grocers record waste for financial reporting and never feed it back into forecasting. It is the closest thing you have to direct evidence of over-ordering by store, SKU and day, and joined to sales and delivery data it turns a one-sided forecast into a two-sided optimisation. Where waste is recorded at all, this is usually the highest-return week of work available.

Markdown, which in fresh is an intraday decision

Time of day matters more than depth

A reduction applied at four in the afternoon clears stock that a deeper reduction at seven will not, because the customers who buy reduced items shop at particular times. The optimisation is over timing and depth together against your own historical clearance response.

Model the effect on full-price demand

Customers who learn that reductions appear reliably at a given hour will wait for them. A markdown policy that clears stock efficiently while training regulars to buy reduced has moved units and lost margin, and it takes months to show up.

The waste alternative sets the floor

For a product that will be waste tomorrow, almost any price above handling cost beats disposal, and the decision changes character in the final hours. Encode that explicitly rather than applying a fixed reduction ladder.

Store execution is the binding constraint

A markdown recommendation nobody has time to apply is a report. Volume of reduction decisions per store per day has to fit the labour available, and that is a design parameter rather than an afterthought.

Online grocery, where substitution decides everything

Online grocery has an unusual property: a meaningful share of orders cannot be fulfilled exactly, so the substitution decision — made by a picker in seconds, at scale — has a large effect on customer retention and almost never gets modelled.

  • A bad substitution costs more than a missing item. Customers forgive an absence and remember a substitution they found absurd. Rejection rate by substitution type is the metric to build on.
  • Personal history beats category rules. This customer's own past acceptances tell you far more than a generic substitution matrix, and most systems use the matrix.
  • Some categories should never substitute. Allergen-relevant items, baby formula and specific dietary products need hard rules rather than a similarity score.
  • Picking route optimisation is ordinary and valuable. Sequencing a pick list against store layout and item characteristics saves real minutes per order.
  • Predict short picks before the picker walks. Knowing at order time which items are likely unavailable lets you offer the customer a choice rather than surprising them at delivery.
Process

How an engagement runs

Waste data joined to sales first, because it turns a one-sided forecast into a real optimisation.

Weeks 1 to 2

Data and constraint review

Waste recording quality, delivery cadence, order deadlines and how many markdown decisions a store can actually execute.

Weeks 3 to 7

Fresh data foundation

EPOS, waste, deliveries, weather and promotions joined at store-SKU-day, with censored demand corrected.

Weeks 8 to 13

Fresh forecasting and markdown

Short-horizon forecasting with waste as a modelled cost, and markdown optimised on timing as well as depth.

Weeks 14 to 18

Store pilot

Against matched control stores, measured on waste and availability together.

Weeks 19 onward

Rollout

By region, with execution volume monitored so recommendations stay actionable.

Deliverables

What you receive

Less waste and better fresh availability, measured against control stores.

01

Waste-aware demand model

Waste joined to sales, so over-ordering is observable rather than inferred.

02

Short-horizon fresh forecasting

Daily store-SKU forecasts with weather and promotional effects modelled.

03

Markdown optimisation

Timing and depth together, with the effect on full-price demand modelled.

04

Execution-constrained recommendations

Sized to the number of decisions a store can actually action.

05

Substitution recommendation

Built on personal acceptance history, with hard rules where substitution is unsafe.

06

Controlled pilot results

Waste and availability against matched control stores.

Fit check

Is this the right starting point?

Worth being direct. There are situations in grocery where custom AI work is the wrong spend, and those are listed rather than buried.

Worth doing if

  • Fresh waste is a material cost and waste data is recorded but never modelled.
  • Markdown is applied on a fixed ladder rather than optimised on timing.
  • Online substitution rejection rates are high and driven by a generic matrix.
  • Fresh availability and waste are managed by different people with opposing targets.
  • You have store-SKU-day data with delivery and weather history attached.

Do something else if

  • Waste is not recorded at store-SKU level. Start there; it is a process change, not a project.
  • Delivery cadence means a better daily forecast changes nothing.
  • Store labour cannot execute additional markdown decisions.
  • The real constraint is supplier lead time or cold chain capacity.
Questions

Frequently asked questions

Marked up with FAQPage schema so these answers can surface directly in search results and inside AI assistant responses.

Where does grocery AI actually pay?

Fresh. Ambient forecasting is ordinary retail forecasting and the vendor tools handle it adequately. Fresh has a shelf life in days, a daily decision cadence, real weather sensitivity and total loss on over-ordering — which means the decisions are frequent, consequential and far beyond what any person can review. It is also where waste data gives you a signal most grocers never use.

We record waste but do not use it. Is that a problem?

It is the single biggest missed opportunity we see in grocery. Waste is direct evidence of over-ordering by store, SKU and day, and joining it to sales and delivery data converts a one-sided forecast — which only sees what you sold — into a genuine two-sided optimisation. Where the recording is decent, this is usually the highest-return week of work available.

How should markdown decisions be made?

On timing and depth together, not on a fixed reduction ladder. A shallower reduction earlier frequently clears more than a deeper one later, because reduced-item shoppers come at particular times. Two things must be modelled alongside it: the effect on full-price demand, since regulars learn to wait for reliable reductions, and the number of decisions a store can physically execute in a day.

Can AI improve online substitution?

Substantially, and it is under-invested relative to its effect on retention. The key is using the customer's own acceptance history rather than a generic category matrix, and treating rejection rate by substitution type as the metric. Some categories — allergen-relevant items, baby formula, specific dietary products — need hard rules rather than similarity scores, and that should be explicit in the design.

Does weather really matter that much?

In specific fresh categories, yes, and considerably more than in ambient. Barbecue lines, salads, ice cream and soft drinks respond sharply and non-linearly, and the response differs by store depending on catchment and format. It is worth modelling properly rather than as a single temperature coefficient — and worth checking that the forecast arrives before the order deadline, because a weather-aware forecast that lands after ordering is a report.

Tell us what the problem looks like.

Thirty minutes, no charge, no deck. We will tell you whether this is an AI problem, a data problem, or a process problem — and we will say when the honest answer is to buy something rather than build it.